Reframing Deep Reinforcement Learning for Large-Scale, Real-World Implementation

Reinforcement learning applied to robotics systems has the potential to revolutionize many industries, but still lacks the flexibility, scalability and safety needed for deployment. One key bottleneck of such methods is their reliance on obtaining information from hand-engineered reward functions, requiring a non-trivial design effort in tasks where experience is expensive to collect.

The goal of this research project is to augment the classical reinforcement learning approach to develop efficient, practical and scalable algorithms for robotics control. The ultimate objective is to produce and deploy a versatile and safe system, able to learn a set of customizable tasks ranging from industrial processes to human support at run time. This entails providing autonomous agents with the ability to recover a structured representation of their environment, obtain hierarchical knowledge over higher-level behaviour and utilize social signals to best understand their objectives. The current plan is to approach each of these sub-goals individually and ultimately implement them jointly on the Toyota HSR Robot. The expected research outcomes are:

- Develop a novel method to enable usage of human interaction to form a meaningful form of information for learning new tasks.

- Define a framework for the unsupervised recovery of a higher-level disentangled representation of the agent's mechanical inputs.

- Incorporate a system for meta-learning, bolstering training efficiency over a range of user-specified tasks.

- Port these systems on the HSR robot and evaluate real-world scenarios of interaction with human users.